01
AI research
Applied research on agents, LLMs, and world models — how they reason, how they fail, and how to make them reliable in real work.
An AI product lab
Building the future of human–agent collaboration.
Herald Labs does applied AI research on agents and turns what we learn into products — figuring out how humans and AI agents will actually work together.
Current surface
Agents, humans, and production pressure.
Research
Product
Research
Agents, LLMs, and world models under real pressure.
Applied research on how models reason, how they fail, and how to make them reliable in real work.
Products
Research that ships as working software.
Every line of research ends in something you can use — an application, a tool, a system.
ModeBuild first
StackHumans + agents
ProofWorking demos
What we do
Research that ships as product.
Herald Labs doesn't publish papers and stop there. Every line of research ends in something you can use — an application, a tool, a working system.
02
Agent applications
Very cool applications built around agents — products where an agent isn't a demo, it's the engine doing the actual work.
03
Human–agent work
New interfaces, workflows, and operating patterns for teams made of humans and agents working side by side.
04
Models, end to end
From frontier LLMs and world models down to the scaffolding, evaluation, and tooling that make raw capability dependable.
05
Field notes
Lessons from agent operations, evaluation, and QA — published so the next person or agent can build on them.
Mission
Figure out agents and human work.
The future of work has agents in it. Our job is to work out what that actually looks like — and build it, one working system at a time.
Operating method
Start with the question. End with proof.
Each project is small enough to ship, concrete enough to test, and documented well enough for the next person or agent to continue without folklore.
Frame the question
Start with a real question about how agents and humans work — not a technology looking for a use case.
Build the thin slice
Humans and AI agents build the smallest end-to-end version that can be honestly judged.
Test against reality
Run it on real work, real documents, real users. If it only works in a showroom, it doesn't work.
Ship the evidence
Release the product, publish the lesson, keep the receipts — then scale it or kill it.
Join us
The future of work has agents in it. We're building it.
If you're working on agents, want to build with us, or have a hard problem that needs one — we want to hear from you.